1
$\begingroup$

I'm exploring follow-the-regularized-leader FTRL proximal gradient descent: paper, reference implementation.

Everywhere FTRL is mentioned, the loss surface for the gradient decent is the LogLoss, and the model for prediction is Logistic regression.

Can I use the same algorithm for a linear least squares model? I have a problem I want to model with a linear model and define the loss by least squares, and then do FTRL to find the optimal solution - do you see any problem with that?

Thanks.

$\endgroup$
1
$\begingroup$

As we can see, FTRL is nothing to do with our model, and it only serves as a way of approximating the optimal solution , just as SGD does.

Logloss is just a kind of loss pattern in classification problem,and there are other alternatives for it , such as ExpLoss. Least-squres loss is one of all loss patterns in regression problem.

The reason why FTRL was mentioned everywhere ,which mainly treated Logloss as a example, is that it was often used in classification problem , such as CTR model in online advertisement scenario.

Hopes this can helps you ,and good luck !

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.